Binary Classifier Ordering for Imbalanced Multi-Class Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Multi-class classification problems are susceptible to bias due to imbalanced training data, making it difficult to build accurate classifiers as the number of classes increases, and existing techniques to address this issue are complex and time-consuming.
Innovation Solution
A multi-class classification system trains a series of binary classifiers in a hierarchy, removing classified data points from the training dataset as they are classified, and automatically determines the order of classifiers to mitigate bias, using techniques like one-vs-all classification and adaptive ordering to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional multi-class classification methods are used, then classification capability is provided, but bias occurs due to imbalanced training data and accuracy decreases as the number of classes increases
Solution Approach 1:
The multi-class classification problem is segmented into multiple binary classification problems. Each binary classifier handles one class at a time, comparing it against all other classes. This segmentation allows each classifier to focus on a specific class, making the training process more effective with imbalanced data and reducing overall bias.
Solution Approach 2:
The approach transforms the multi-class classification problem into a series of binary classification problems by adding a dimensional transformation. Instead of directly classifying among multiple classes, the system creates binary decisions for each class, effectively changing the problem dimensionality from multi-class to multiple binary problems.
2Measurement precision
If existing techniques are used to address bias in multi-class classification, then some accuracy improvement may be achieved, but the training process becomes complex and time-consuming
Solution Approach 1:
The training process is segmented into independent binary classification training tasks. Each binary classifier is trained separately on its corresponding class, which simplifies the overall training complexity compared to training a single multi-class classifier on imbalanced data. This segmentation makes the training process more manageable and less time-consuming.
3Measurement precision
If a series of binary classifiers are arranged in different orders, then training results vary, but determining the optimal order increases computational complexity
Solution Approach 1:
The system dynamically determines the arrangement order of binary classifiers based on the specific characteristics of the training data. Rather than using a fixed or random order, the arrangement is adapted to the data distribution, which helps achieve more consistent training results while avoiding the need to evaluate all possible permutations.
Data Source
AI summary
An embodiment generates a training batch of data points from training data for a plurality of classes and builds a multi-class classifier having a series of binary classifiers arranged in a first order. Each of the binary classifiers is associated with a respective class. The embodiment trains the multi-class classifier with the binary classifiers arranged in a first order and, at each binary classifier, the embodiment identifies data points as belonging to the class associated with the respective classifier and updates the training batch to exclude the classified data points. The embodiment then modifies the multi-class classifier by changing the order of classifiers and repeats the training of the multi-class classifier with the series of binary classifiers arranged in a second order. The embodiment then selects a final configuration of the multi-class classifier based at least in part on a comparison of first training results to the second training results.